Language
description
Transcript of Language
Language
Richard L. Lewis Psychology,
Linguistics, and Electrical Engineering & Computer Science
University of Michigan
July 24, 2001
ACT-R and Language Processing:Opportunities, Challenges, and the Linguistic Killer Bees
Richard L. Lewis Psychology,
Linguistics, and Electrical Engineering & Computer Science
University of Michigan
July 24, 2001
Wexler’s (1978) review of Language, Memory, & Thought Wexler’s conclusions:
• “There is not much prospect of adding scientific knowledge by pursuing the methods … in LMT”
• “There is remarkably little that a linguist (or even psychologist) could learn by reading LMT”
Some major gripes • Trying to achieve too much with a broad theory
– Nature of explanation may be different in different domains (modularity)
• Theory not “restrictive”—no explanatory principles– ACT can model anything, therefore explains nothing
• Not dealing with sufficient range of complex linguistic phenomena
Wexler’s diagnosis and cure Core problem: representational weakness “In studying language processing it seems
obvious that the use of a strong representational theory would be very helpful...In particular, such a theory exists for syntax.”
Work toward theory of “performability”• How can structures uncovered by linguistic theory be
processed by a processor with human constraints?• Important that “entire representational theory be
processed”
What can ACT contribute to psycholinguistics and linguistics?
Good news, and a question
On the prospects for a productive science of cognition grounded in cognitive architectures (& ACT), Wexler was simply wrong• The present state-of-the-art in architectural
theorizing (as represented in ACT-R) is evidence of significant progress over last 20+ years
But for language, we can still ask:
By the way, John is in good company From Bickerton’s (1992) critique of Newell’s
(1990) UTC chapter on language:• “To criticize the eighteen pages on language and
language acquisition that follow would be like shooting fish in a barrel”
Part of Newell’s response:• “… I clearly faltered in dealing with the linguistic
killer bees…. The lesson, of course, is never, never show fear.”
Overview(1) Preliminary: Major choices in developing ACT-R
models of language(2) An example domain: sentence processing
• Brief sketch of the structure of one model • Some interesting—even unexpected—theoretical and
empirical issues very closely tied to architecture
(3) What does ACT-R buy you? Major opportunities for ACT-R in language
(4) Potentially serious challenges(5) Revisiting the killer bees
A major choice in developing language theories in ACT-R Should linguistic processes be realized within
ACT-R, or should ACT-R be treated as central cognition, and a language module developed for it?
Perhaps we should work on an ACT-R/PML• Where the “slash” corresponds to Jerry Fodor’s way
of “carving the mind at the joints That’s not the approach I’ll discuss today—nor is
it the approach traditionally taken in ACT research
The alternative
Treat language as a cognitive skill; embed linguistic processing in the architecture
We know from Soar work that this can (surprisingly) yield processing models that are consistent in many respects with modular approaches
And, even if it turns out to be wrong, we need to know why
Another major choice How to distribute lexical and grammatical
knowledge across declarative and procedural memory
Approach I’ll assume:• Lexicon in declarative memory• Grammatical knowledge in procedural memory (for
comprehension, in form of parsing productions)• Field typically doesn’t phrase distinction in these terms,
but one notable exception: Consistent with at least one neuropsychologically motivated model (the “declarative-procedural model”, M. Ulmann)
Why sentence processing?
Because that’s mostly what I work on
Also: A very interesting combination of symbolic and subsymbolic, and fast, real-time but complex processing
Incredibly rich empirically and theoretically (perhaps too much so)
A classic processing limitation Most people find one level of embedded clause
comprehensible:
The dog that the cat chased ran away.
But double center embeddings are very difficult (Miller & Chomsky 1963):
The salmon that the man that the dog chased smoked tasted bad.
WHY?
Examples of good/simple ideas that don’t quite work
Kimball's Principle of Two Sentences: Can't parse more than two sentences at once• [s What [s the woman that [s John married] likes] is
smoked salmon.] (Cowper 1976; Gibson 1991)
Limited buffer for holding unattached NPs (say, two)• John-wa Bill-ni Mary-ga Sue-ni Bob-o syookai sita
to it-ta. (Lewis, 1993, 1996)
• John Bill Mary Sue Bob introduced said
NPSue-ni
NPMary-ga
NPBill-
ni
NPBob-o
NPJohn-wa
Subject?
Ind-Object? introduced
Parsing as associative, cue-based retrieval from WM Construes attachments as associative WM retrievals (Lewis 1998; McElree,
1998, 2000)– Interesting connections to MacWhinney’s (1989) cue-based competition model (& other constraint-based
approaches)
Cues include (at least) syntactic relations• Retrieval interference arises from cue-overlap (a kind of similarity-based
interference)
Decay, focus, & interference
i j
Focused elements serve as retrieval cues. Memory elements receive additional activation from associated focus elements.
Ai Bi WjSjij S ji
Example of activation dynamic in parsing
The boy with the dog
(1.2)
The boy with the dog saw
(3.8)
(3.9)
The boy with
Four interesting theoretical issues(1) How working memory limitations in parsing
arise(2) The representation of serial order
information in sentence processing(3) Modularity and control structure(4) Decay vs. interference in processing
ambiguities
Left alone constituents decay More cues less activation for each More constituents associated with a cue
less effective the cue is
Worst case: multiple distal attachments with high retrieval interference
Issue #1: Implications for WM limitations in parsing
Distance vs. interference Worst case for parser:
• Multiple [limited focus], distal [decay] attachments, with multiple similar candidates [interference]
The boy who the dog that the fish saw ate was
(0.67)
The girl with the dog with the boy with the fish with the ticket was….
(3.7)
…..But long-distance attachments still possible
Contrasts in center-embedding: SR vs. OR
The boy who the dog bit
The boy who the dog bit saw The boy who bit the dog saw
The boy who bit
(2.64)
Subject Relative
(3.45)
(1.34)
Object Relative
(3.08)
Effects of locality: RC/SC contrasts (Gibson, 1998)
The claim that the boy who the dog bit died upset me
(2.57) (1.69)
The man who the claim that the dog bit the boy upset ...
(0.77)
Mary saw the dog.At “saw”, attempt attachment of the subject (not the object). No need to distinguish the relative order of Mary and dog.
Keeping track of serial order information is functionally required
But very often, simply distinguishing current item from preceding items is sufficient• E.g., lets you distinguish the S from the O in SVO
languages
Issue #2: Serial order
When distinguishing current from preceding is not enough General case: When two or more preceding
items must be discriminated solely by their serial positions
Examples: Japanese sentential embeddings:Mary-ga Tom-ga butler-o killed knows
Who killed the butler?
X1 . . . X2 . . . Y?chicken egg
Just attach to the most active (recent) candidate (simulate a stack)
Two functional problems:(1) The activation/strength of an item may not accurately reflect its serial position
– Evidence from STM paradigms suggests item strength is not a good surrogate for position (e.g., McElree & Dosher, 1993)
– Also, some items may be linguistically focused, or may have received additional processing
– So this is a dicey theoretical path
(2) Sometimes it is necessary to attach to the first item, not the most recent
Why not use activation decay?
Cross-serial dependencies In Dutch, a standard embedded construction
requires crossed, not nested, dependencies
• …omdat ik Cecilia de nijlpaarden zag voerenbecause I Cecilia the hippopotamuses saw feed“because I saw Cecilia feed the hippopotamuses. (Steedman, 2000)
NP1 NP2 NP3 V1 V2 V3
How should we represent serial order in such a parser? Major choice: position codes vs. associative
chaining• Relevant STM phenomenon: items in nearby positions
tend to be confused (e.g., Estes (1972))
• Confusable position codes adopted by many researchers (eg., Hensen (1998); Burgess & Hitch (1999), Anderson et al, (1998))
Idea: associate each item with a position code that is a value along some gradient• There is a distinguished START anchor code
(say,START = 1.0); other positions defined as some (probably non-linear) function
Using position codes as retrieval cues in parsing Only two codes are used as cues: START and END
(the current position)• To effect a recency/stack discipline for nested
dependencies, use the END code as a cue• To effect a primacy/queue discipline for crossed
dependencies, use the START code as a cue• The best matching (closest) item will be retrieved (all
other things being equal) No need to assume parser knows about any other
position codes (e.g., the 2nd, or 4th)• I.e., these are not grammatically meaningful
Using position cue to retrieve most recent candidate
NP-ga NP-ni [NP-ga NP-o V] V
1.0 0.78 0.65 0.58 0.49
Retrieval cues: END (= 0.49), Subject
Given:• Confusable positional codes• And the functional requirement to distinguish items based on
position Then it may be possible to make processing easier by
increasing distance—increasing positional distinctiveness
Data from Uehara (1997) bear this out:
• NP-ga NP-ga NP-ga NP-o V V V (4.31)• NP-ga NP-ga Adv NP-ga NP-o V V V (3.61)
Prediction of positional similarity effects
Testing positional similarity with single embeddings 2 x 2 design varying stacking (3 and 4 NPs) and
positional similarity of subject NPs(a) (ps=0, stack=3)
NP-ga NP-ni NP-ga V V my brother teacher girl playing notified
(b) (ps=2, stack=3) NP-ga NP-ga NP-o V Vdentist president interpreter called remembered
(c) (ps=0, stack=4) NP-ga NP-ni NP-ga NP-o V Vprofessor president representative student examined promised
(d) (ps=2, stack=4) NP-ga NP-ga NP-ni NP-o V Vstudent lecturer reporter author introduced noticed
Difficulty rating study
Participants rated difficulty on 7 point scale (1=easy, 7= difficult) • Each participant saw four versions of each
experimental type (16 total experimental sentences interspersed with 34 fillers; 50 total)
• Familiarity of lexical items controlled across conditions
Participants were 60 female students from Kobe Shoin Women’s University in Japan
ResultsMean difficulty rating
(w/ 95% CI)
0123456
Positional similarity
Difficulty rating
Stack 3Stack 4
Stack 3 3.02 5.07Stack 4 4.2 5.22
0 2
It’s robust
Effect shows up in 2 rating paradigms: 1-7 fixed scale, and magnitude estimation
Effect shows up in 3 presentation paradigms• Paper & pencil questionnaire• Self-paced moving window• Self-paced central presentation
The ACT model on the single embeddings
(1.07)(1.62) (2.53)
(5.12)
NP-ga NP-ga NP-ni NP-o V V
(1.08)(1.85)
(2.54)
(4.5)
NP-ga NP-ni NP-ga NP-o V V
Why Dutch is easier than German Consider the positional mismatch when using
START vs. END (current) position codes
Retrieval cue: END
NP1 NP2 NP3 V3 V2 V1
Retrieval cue: START
NP1 NP2 NP3 V1 V2 V3(0)
Perfect positional match
Summary of qualitative coverage
Classic difficult double relatives (French, Spanish, German)
Subject vs. object relatives
Subject sentences with relative clauses
RC extrapositions (German)
Many stacking contrasts (Japanese, Korean)
Cross-serial vs. nested contrast (Dutch, German)
RC/SC vs. SC/RC contrast
Positional similarity contrasts (Japanese)
Various pseudo-cleft and it-cleft w/relative contrasts
Various Wh-movement w/ relative constructions
Issue #3: Modularity and control structure Does ACT-R yield a modular or interactive
account of sentence processing?• The answer may surprise you • [Answers were surprising in Soar as well; see
Lewis, 1998; Newell, 1990] Where to look: Factors affecting on-line
structural ambiguity resolution
Quick review
Example:• Mary forgot her husband
would ……. Structural accounts like
Minimal Attachment prefer the direct object structure over the sentential complement• Because the direct object
structure is less complex (fewer nodes); hence should be computed faster
S
NPMary
VP
Vforgot
NPher husband
S
NPMary
VP
Vforgot
NPher husband
S
Structural ambiguity resolution in the ACT-R model Choice between competing parsing
productions corresponds to selection of a path to pursue in the parsing search space
Thus, structural ambiguity resolution happens via conflict resolution
I.e., the theory of ambiguity resolution is ACT-R’s theory of conflict resolution
Modularity and conflict resolution in ACT-R 3.0 vs. 4.0 Key issue in modularity: are there architectural
boundaries that prevent certain kinds of information from being brought to bear on some processing decision?
In ACT-R 4.0, there is a clear move in the direction of limiting the information flow, compared to 3.0, 2.0• Fewer factors (fewer knowledge sources) affect initial
production choice in ACT-R 4.0 This has a fairly dramatic effect on the nature of the
resulting sentence processing theory
ACT-R 3.0 vs. ACT-R 4.0
ACT-R 3.0: Particular attachments (production instantiations) compete• Activations of declarative elements are a factor in the
conflict resolution• Provides way to integrate decay/recency, and frequency
effects into ambiguity resolution ACT-R 4.0: First, attachment types compete
• Based on their expected gain• Then, given an attachment type, different instantiations
of that attachment compete– I.e., different chunks “compete” for retrieval
An old favorite...The horse raced past the barn fell.
At raced, main verb attachment production will win the initial competition• A much more successful construction; and predicted cost of
reduced relative is higher But considerable evidence now that GP effect can be
reduced (even eliminated?) by various lexical, contextual, semantic factors
The students taught by the Berlitz method failed miserably.
An interesting asymmetry The evidence in favor of on-line semantic/contextual
effects always shows how various factors make the dispreferred structure easier
But Frazier (1995, 1998) has argued that it is “completely accidental in the constraint-satisfaction model that garden paths have not been demonstrated for the simpler (preferred) structure”
The children pushed quickly by the armed guards.
The predictions Constraint-satisfaction models predict:
• Effects of semantic fit,context on processing dispreferred structure (could reduce garden path)
• Effects of semantic fit, context on processing preferred structure– because evidence for the dispreferred structure is evidence against the preferred
structure– So could actually cause a garden path if dispreferred structure becomes
preferred
Classic structural models predict:• Effects of semantic fit, context on processing dispreferred
structure due to easier reanalysis (second pass processing)• But NO effect on preferred structure, because it is always
pursued first
The ACT-R 4.0/5.0 prediction
The ACT-R predictions should pattern with the classic structural/Minimal Attachment-style theories
Binder, Duffy & Rayner (2001) This is exactly what Binder et al found, in a carefully
done eye-movement study They found no hint of garden path effects in first-
pass measures for the main verb construction, even when both semantic fit and referential context conspired against the main verb reading, and for the reduced relative reading• Importantly, they used materials (and even improved on
them) that have been demonstrated to show clear effects of GP reduction for the relative clause, and showed in off-line norming that they were equi-biased
Another unexpected asymmetry Competition between structure types should
not show effects of recency (decay)• E.g., VP-PP attachment vs. NP-PP attachment• Because the conflict resolution won’t consider the
activation of the attachment site But competition between sites for the same
type of structural attachment should show effects of site activation (perhaps recency)• E.g., VP1-PP vs. VP2-PP
Example of the asymmetry in PP attachment Between structure-types: VP vs. NP
• Mary painted the wall with cracks.• Competing sites are the VP painted and the NP wall
Within structure-type: NP1 vs. NP2• The father of the queen with the beard.
This actually maps roughly onto distinction between two major preferences: Minimal Attachment and Late Closure (Frazier)• BUT: It has always been stipulated that when the two
factors conflict, MA wins
Bottom line theoretical implications ACT-R 3.0 yields something a bit closer to
lexical-constraint-based approach to ambiguity resolution (e.g., Tanenhaus, MacDonald…)
ACT-R 4.0/5.0 yields something a bit closer to a modular structure-first approach to ambiguity resolution (e.g., Frazier, Clifton)• Actually, closest to statistical tuning models (e.g,
Mitchell et al, Crocker et al.) WARNING: I’m oversimplifying the issues
here considerably
Example of possible problems: Major category frequencyBoland 1998; Corley & Crocker 1998 Base activation of lexical entries reflects
frequency; determines retrieval latencies Ambiguous: bias affects resolution
• the German makes the beer/are cheaper… the warehouse prices the beer /are cheaper
• All things being equal, base-level activations will determine which lexical entry is attached first
Unambiguous: bias affects processing times• Lower base-levels = slower times for subordinate• …the German make is cheaper than ..
Factoring into multiple productions may save the day A production could make an initial retrieval of the
dominant lexical entry, followed by the attachment productions
More consistent with smaller-grain-sized productions anyway
Something like this happens already in the ACT-R 3.0 model• But these productions are category-specific; this
solution will only work if productions are general Critical issue, then, may be TIME
Issue #4: Decay vs. interference in reanalysis Any serial model (such as the ACT-R model) of
sentence processing must be capable of reanalysis when the wrong path is pursued• Because not all “garden paths” are difficult:
The boy understood the man was paranoid. What factors affect reanalysis difficulty?
• One common assumption is length; but even long ambiguous regions can be pretty easy:
The boy understood the man who was swimming near the dock was paranoid.
Another interesting asymmetry prediction Both structural interference (due to associative
retrieval interference) and decay should affect syntactic attachments• Hence, should also affect reanalysis; because reanalysis
requires attachment to the correct (dispreferred, discarded) structure
But a discarded structure will suffer MORE decay than the chosen structure, because the chosen structure receives activation boosts from being used
But it will NOT suffer more interference
A reanalysis study(w/ Julie Van Dyke, Pitt/UM)
Compare ambiguous and unambiguous versions of short, long, and interfering structures• We compute cost of reanalysis per se by comparing
ambiguous and unambiguous conditions Example of interfering condition:
• The boy understood the man who said the townspeople were dangerous was paranoid.
Task: rapid grammaticality judgment (Ferreira & Henderson, 1991)
The results
0.4
0.5
0.6
0.7
0.8
0.9
1
Unambiguous Ambiguous
SENTENCE TYPE
PER
CEN
TAG
E C
OR
REC
TShort Interfering Long
Results replotted
-0.3
-0.2
-0.1
0
0.1
ATTACHMENT REANALYSIS
AC
CU
RA
CY
DIF
FER
ENC
E
DISTANCE INTERFERENCE
Some distinctive features of the ACT-R approach Explicit theory of retrieval in WM
• Fits well with emerging modern views of STM/WM (McElree, Tehan & Humphreys, …)
• And effects of decay, interference Explicit theory of serial order representation
• Tackles long-neglected functional problem• Potential unification with verbal STM theory
New perspective on modularity, grounded in rational analysis and computational concerns• “Independently motivated” theory (ACT-R conflict resolution)
providing great constraint Generally, contact with cog psych theory
What does ACT-R buy you?(1) ACT-R is a vehicle for making psycholinguistics come into
contact with the theoretical vocabulary of cognitive psychology
(2) Evidence in some areas that the details of ACT-R may be pushing the theory in just the right direction
(3) Provides unification with cognitive theory; hence greater explanatory power
(4) Provides framework for building detailed quantitative processing models• Hence, permits bringing to bear quantitative data on theory
construction
A new kind of psycholinguistics? ACT-R modeling, and the empirical work it
motivates, could help lead to a psycholinguistics characterized by• Complete, detailed processing theories; models of the
fine structure of sentence processing• Quantitative, parameter-free(?) models• Models that make explicit predictions about the
dependent measures used in the experiment (eye-movements, button presses, judgments)
• Highly constrained, hence explanatory and predictive models
Major opportunities and challenges(1) Incorporating independently motivated ACT
language models in all models involving verbal material.
(2) Instruction taking.(3) Functional NLP concerns: Scaling up(4) Linguistic task operators(5) Closing the perception-motor loop via ACT-R/PM
All are unique to cognitive architectures
Opportunity/challenge #1
Routinely incorporating independently motivated ACT language model(s) in all models of experiments with verbal materials• Closing the loop so that the linguistic processing is
completely constrained; no theoretical degrees of freedom on the language side (cf. Kintsch models of comprehension)
• Some examples moving in this direction: Anderson, Budiu, & Reder (2001), and Altmann & Davidson (2000)
• Verbal rehearsal
Opportunity/challenge #2
Instruction taking (one of Newell’s dreams for a UTC)• Finally “close the loop”: rather than posit
productions and chunks that encode knowledge of a task, have models that read instructions and carry out the task (Lewis, Polk, Newell, 1989)
• Considerably reduces theoretical degrees of freedom
• Build systems that accomplish variants on some experimental paradigm
Opportunity/challenge #3 Linguistic task operators (Lehman, Polk, Newell,
Lewis, 1991)• Build models in which language is used to perform
cognitive tasks (thinking by talking to oneself)• Uses language comprehension operators themselves as
the interpretive process that yields “behavior” Turns standard instruction-taking process on its
head• Uses NL itself as the language for representing behaviors• Newell had produced a set of LTOs that accomplished the
blocks world
Some speculations Could this offer solution to John’s ugly interpretive code? Depends on an NL semantics very closely grounded in
perceptual-motor representations• So that one’s understanding of “push the button” is quite close
to the motor program that will be set up to actually push the button
• Then “interpretive execution” is more like “releasing” the motor program rather than “interpreting a declarative representation”
Need to be careful: could lead to a procedural semantics to NL• Learn from Miller & Johnson-Laird’s program
Opportunity/challenge #4
Functional NLP challenges• Scaling NL systems in ACT-R: Can ACT handle a
lexicon of 30,000 words? A grammar base of 1,000s of production rules?
• “Train” ACT-R on large corpora of text to set production/declarative memory parameters
• Not just a technical engineering question; of critical theoretical importance psycholinguistically
• “Scale counts in cognition”
Opportunity/challenge #5 Use ACT-R/PM/EMMA to develop explicit models of
eye-movements in reading• Good for the psycholinguistic theory• Good for ACT-R/PM—enormous literature on eye
movements in reading Develop models of eye-movements in context Develop models of button-pressing paradigms
• Word-by-word reading Develop explicit theories of global judgments
(grammaticality, difficulty, acceptability)• Binary, 1-7 scale, magnitude estimation
What does it take to meet these challenges? Common requirements for many:
• Incorporating a (broad coverage) theory of semantic representations
– Possibilities include: Sowa’s Conceptual Graphs; Jackendoff’s Conceptual Structures; Miller et al’s WordNet
• Incorporating a (fairly broad coverage) theory of syntactic and lexical representations
– Possibilities include: HPSG, combinatory categorial grammar
Technically, this incorporation will involve bringing some existing large database into ACT-R
Potentially serious architectural challenges Time—is there enough?
• Probably not.• Only have 200-300ms/word (on average)—Time for a
couple of productions and retrievals• Anderson et al (2001) met constraint by combining
considerable amount of syntactic, semantic structure building into single productions
• But I have separated these in my model (and in earlier NL-Soar model, in which there was just barely enough time)
• Also referential processing happening on-line I don’t taking timing seriously in current model
Potentially serious architectural challenges Is ACT-R too hopelessly symbolic and serial for
language processing?• Many think of “lexical entries” and “lexical access/retrieval” as
old-fashioned Right approach: Go after signature phenomena
addressed by connectionists• Good candidate: Tanenhaus et al work on eye-movements in
context that track time-course of lexical access and sentence processing
• Can see neighborhood effects on-line, extremely rapid match to referential context
• Can ACT-R work fast enough to do this?
Potentially serious architectural challenges Acquisition
• Hard to work on this problem without a stable production learning mechanism
• Perhaps compilation will be a reasonable base Control structure
• Can ACT-R’s conflict resolution handle interactive/lexical effects in ambiguity resolution?
• (Perhaps, but in appropriate time limits?)
What about ACT-R 5.0? By gosh, it’s the best thing since ACT-R 4.0!! A big potential win as I see it now: competitive
declarative memory retrievals• Could provide natural account of differential pattern of
reading times on lexical ambiguities (slow-down) and syntactic ambiguities (no-effect or even faster)
• Could provide less heavy-handed approach to getting associative interference effects; e.g., may not have to worry about dynamically resetting cues so fan doesn’t build up too much
• Potentially cleaner account of similarity-based interference– Probably incorporating Raluca’s representational similarity
Other potential wins in 5.0
Another potential win: parallel retrieval and production firing• Issue: may not be enough time for firing syntactic,
semantic, referential processing productions AND perform lexical access
• But there MIGHT BE enough time if lexical access for the next word can be initiated while finishing up processing of the last
• Predicts “spill-over” effects in reading, for which there is ample evidence, in both self-paced reading and eye-tracking
Timing in ACT 3/4 vs. 5.0
ACT 3.0/4.0
ACT 5.0
Lexical access Syntactic Semantic Referential
Lexical access
Syntactic Semantic Referential
Lexical access
Semantic Referential
Yet another possibility
Use production compilation to compile out the lexical access• Produces word-specific comprehension production
– Similar to original idea of “comprehension operators” in Soar
• Would then shift burden of lexical frequency effects to procedural memory
Has interesting effect of distributing (redundantly in quite specific ways) grammatical knowledge across the “lexicon”• This might be exactly right
Actually, may be impossible to avoid
Revisiting the killer bees Let’s reconsider some of Wexler’s gripes
• Tackling broad range of complex linguistic (and psycholinguistic) phenomena? YES
• Constrained, explanatory principles? YES• Asking for too much from a broad theory (because modularity is
right)? NO And his suggestion of taking advantage of existing
representation theories• This is basically right on target and exactly what was done in
NL-Soar, and is right path for ACT-R• Linguistic theory provides the ontology of representational
features, ACT-R the processing architecture
Minimal Attachment review Dominant theory in the 80’s, early 90’s:
Frazier’s Garden Path Model • Serial (single structure pursued)• Decision principle: Minimal Attachment: Structural
ambiguities resolved by pursuing the simplest structure (determined by counting number of syntactic nodes); simplest structure assumed to be computed most quickly
• (Other principles involved; e.g., Minimal Chain Principle, Right Association; Construal)
Opportunities & challenges, cont. Connectionism
• Will the symbolic side of ACT-R language models be hopelessly symbolic?
• E.g., many researchers reject the idea of “retrieving entries” from a lexicon—stored lexical entries are old-fashioned